Skip to content

DeepSeek R1 Distill Llama 70B VRAM requirements

How much VRAM DeepSeek R1 Distill Llama 70B needs
4-bit weights: 39.8 GBWith 8K context: 43.3 GBWith 32K context: 50.8 GB8-bit, 8K context: 73.3 GB
Parameters
70.6B
Longest context
128K
Released
Jan 2025
License
MIT
Strongest at
Not rated

How much VRAM DeepSeek R1 Distill Llama 70B needs

Memory DeepSeek R1 Distill Llama 70B needs at each precision and context length
PrecisionWeightsWith 8KWith 32KWith 128K
16-bit131 GB135 GB142 GB172 GB
8-bit69.8 GB73.3 GB80.8 GB111 GB
6-bit53.9 GB57.3 GB64.8 GB94.8 GB
5-bit46.7 GB50.2 GB57.7 GB87.7 GB
4-bit39.8 GB43.3 GB50.8 GB80.8 GB
3-bit32.1 GB35.6 GB43.1 GB73.1 GB
2-bit24.6 GB28.1 GB35.6 GB65.6 GB
1.6-bit13.6 GB17.0 GB24.5 GB54.5 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run DeepSeek R1 Distill Llama 70B

The best way DeepSeek R1 Distill Llama 70B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4060 report
RTX 5060Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 5060 report
RTX 3070Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 3070 report
RTX 3080 10 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 3080 10 GB report
RTX 3060 12 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 3060 12 GB report
Arc B580Too bigPDF: DeepSeek R1 Distill Llama 70B on Arc B580 report
RTX 4070Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4070 report
RTX 4070 SuperToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4070 Super report
RTX 5070Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 5070 report
RTX 4060 Ti 16 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 5060 Ti 16 GB report
Arc A770 16 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on Arc A770 16 GB report
Radeon RX 9070 XTToo bigPDF: DeepSeek R1 Distill Llama 70B on Radeon RX 9070 XT report
RTX 4070 Ti SuperToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4070 Ti Super report
RTX 4080 SuperToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4080 Super report
RTX 5070 TiToo bigPDF: DeepSeek R1 Distill Llama 70B on RTX 5070 Ti report
RTX 5080Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 5080 report
Radeon RX 7900 XTXToo bigPDF: DeepSeek R1 Distill Llama 70B on Radeon RX 7900 XTX report
RTX 3090Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 3090 report
RTX 4090Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 4090 report
Radeon AI PRO R9700Too bigPDF: DeepSeek R1 Distill Llama 70B on Radeon AI PRO R9700 report
RTX 5090Too bigPDF: DeepSeek R1 Distill Llama 70B on RTX 5090 report
2× RTX 30903-bit36K18 tok/s2.4 sPDF: DeepSeek R1 Distill Llama 70B on 2× RTX 3090 report
2× RTX 40903-bit36K20 tok/s1.0 sPDF: DeepSeek R1 Distill Llama 70B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB3-bit36K26 tok/s2.2 sPDF: DeepSeek R1 Distill Llama 70B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada3-bit36K19 tok/s1.6 sPDF: DeepSeek R1 Distill Llama 70B on RTX 6000 Ada report
L40S3-bit36K17 tok/s2.6 sPDF: DeepSeek R1 Distill Llama 70B on L40S report
2× RTX 50906-bit16K21 tok/s0.8 sPDF: DeepSeek R1 Distill Llama 70B on 2× RTX 5090 report
4× RTX 30908-bit56K8.7 tok/s1.3 sPDF: DeepSeek R1 Distill Llama 70B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit64K16 tok/s1.1 sPDF: DeepSeek R1 Distill Llama 70B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on Mac mini M4 16 GB report
Mac mini M4 24 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on Mac mini M4 24 GB report
Mac mini M4 32 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GBToo bigPDF: DeepSeek R1 Distill Llama 70B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB4-bit16K5.1 tok/s48 sPDF: DeepSeek R1 Distill Llama 70B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit64K5.4 tok/s24 sPDF: DeepSeek R1 Distill Llama 70B on M4 Max 128 GB report
M5 Max 128 GB8-bit64K6.0 tok/s6.9 sPDF: DeepSeek R1 Distill Llama 70B on M5 Max 128 GB report
M3 Ultra 96 GB6-bit40K8.8 tok/s13 sPDF: DeepSeek R1 Distill Llama 70B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit128K6.9 tok/s13 sPDF: DeepSeek R1 Distill Llama 70B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit128K6.9 tok/s13 sPDF: DeepSeek R1 Distill Llama 70B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit128K9.9 tok/s4.2 sPDF: DeepSeek R1 Distill Llama 70B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB4-bit16K3.6 tok/s14 sPDF: DeepSeek R1 Distill Llama 70B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit64K2.1 tok/s14 sPDF: DeepSeek R1 Distill Llama 70B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit128K2.8 tok/s4.8 sPDF: DeepSeek R1 Distill Llama 70B on RTX Spark report
DGX Spark8-bit128K2.6 tok/s4.8 sPDF: DeepSeek R1 Distill Llama 70B on DGX Spark report
Dual DGX Spark8-bit128K2.6 tok/s2.8 sPDF: DeepSeek R1 Distill Llama 70B on Dual DGX Spark report
DGX Station GB3008-bit128K58 tok/s0.4 sPDF: DeepSeek R1 Distill Llama 70B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit128K16 tok/s0.7 sPDF: DeepSeek R1 Distill Llama 70B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit128K16 tok/s0.4 sPDF: DeepSeek R1 Distill Llama 70B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit128K16 tok/s0.2 sPDF: DeepSeek R1 Distill Llama 70B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit16K18 tok/s3.1 sPDF: DeepSeek R1 Distill Llama 70B on A100 80 GB SXM report
H100 SXM8-bit16K30 tok/s1.0 sPDF: DeepSeek R1 Distill Llama 70B on H100 SXM report
4× H200 NVL8-bit128K41 tok/s0.4 sPDF: DeepSeek R1 Distill Llama 70B on 4× H200 NVL report
8× H200 NVL8-bit128K41 tok/s0.2 sPDF: DeepSeek R1 Distill Llama 70B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF
vLLM
vllm serve deepseek-ai/DeepSeek-R1-Distill-Llama-70B
MLX (Mac)
mlx_lm.chat --model mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit

Questions

How much VRAM does DeepSeek R1 Distill Llama 70B need?

DeepSeek R1 Distill Llama 70B needs 39.8 GB for its weights at 4-bit, 43.3 GB with an 8K context and 50.8 GB with 32K. At 8-bit, 73.3 GB.

What is the minimum VRAM for DeepSeek R1 Distill Llama 70B?

About 35.6 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the 2× RTX 3090.

How big is DeepSeek R1 Distill Llama 70B?

70.6B parameters: 131 GB of weights at 16-bit, 39.8 GB at 4-bit.

Can DeepSeek R1 Distill Llama 70B run on an RTX 4090?

Not on one card. The smallest machine here that runs it is the 2× RTX 3090.

More models